Re: GTK+-2.x planning

[Owen Taylor <otaylor redhat com>]

Daniel Egger <degger fhm edu> writes:

> Am Don, 2002-03-14 um 22.08 schrieb Owen Taylor:
> 
> > The two important optimizations to be made in gdk-pixbuf/pixops at this point
> > are:
>  
> >  A) Special-case compositing/copying with no scaling. 
> 
> Please explain or point to a bug.

Pretty simple ... gdk_pixbuf_composite_* and gdk_pixbuf_scale always run through
the scaling code path, even when the scale factor in both directions is 1...
for this common case, presumbably big wins could be found by adding a 
non-scaling code path.
 
> >  B) Fix the scaling-down-by-a-large-factor problem.
> 
> > Especially B). 
> 
> I'll look deeper into this when I solved the automake 1.4/gcc
> -save-temps problem which is quite nasty for me.
> 
> > Altivec, more MMX, etc, is marginal compared to these, though certainly
> > more fun to hack on ;-)
> 
> I'm hunting bugs for the gcc people right now, but 3.1 will be the first
> release to fully support AltiVec and there are also new builtins to ease
> the use of MMX/SSE without the need to use inline assembly or nasty
> linker hacks to get the stuff going.
> 
> What this means is:
> I'd like to check via configure.in for a gcc >= 3.1 and if it is there
> use the new builtins to enhance MMX/SSE/AltiVec support a fair bit. This
> leads to better code because the scheduling will be much more aggressive
> than can ever be with mixed inline assembly while it is much easier to
> code and fix. Obviously before this bottlenecks have to be identified
> and code parallelised to make this possible.
> 
> My question is: when (in terms of the new schedule) is the right point
> to think about these steps?

 - After GCC 3.1 is released 
 - After you have some prototypes; that is, I see no particular reason
   to the early development of this stuff in CVS.

So, presumably after 2.2. Certainly altivec pixops code could go in
2.4 if it was available and tested at that point. I think ditching the
current MMX in favor of GCC builtins entirely might want to wait a
bit...  at least until 3.1 becomes widespread.

Regards,
                                        Owen